OpenAI launches GPT-6 Astra, a more powerful and autonomous AI model.

OpenAI has officially launched GPT-6 Astra, its new artificial intelligence model described as a "generational leap" toward more autonomous systems. President Greg Brockman explicitly spoke of entering the era of artificial general intelligence (AGI).

The model, unveiled on September 4, 2026, promises advanced capabilities: it can autonomously use a computer, browse online, code, analyze data, and identify problems without explicit instructions. In benchmarks, Astra shows significant improvements, for example achieving 72.6% on the OSWorld 2.0 test and completing tasks in approximately 47% less time compared to its predecessor GPT-5.6 Sol.

Deployment has already begun for a select number of organizations and, within a few days, will be extended to users of ChatGPT Plus, Pro, Business, and Enterprise plans. The model will also be available via the OpenAI API and on AWS.

GPT-6 Astra represents a crucial evolution for OpenAI, aiming to reclaim leadership in the sector after a year in which competitors like Anthropic have gained ground, particularly in coding and among enterprises. The industry challenge is shifting from having the model with the best benchmarks to building an agent capable of working over long periods, making fewer errors, and reducing costs per completed task.

The most delicate front of the launch is cybersecurity. Astra is the first OpenAI model to surpass the "Critical" threshold in the company's internal Preparedness Framework, its risk assessment grid. During testing, it discovered two previously unknown zero-day vulnerabilities. To manage these risks, OpenAI is strengthening protections and launching the "Daybreak" program, offering limited early access to security professionals for vulnerability testing.

The model arrives at a time when the previous GPT-5.6 has just completed its GDPR-compliant deployment in Europe. Astra, at launch, does not yet have a European Data Zone, raising questions about immediate access for users in the old continent. Its efficiency promises to lower estimated API costs by about 57% for complex tasks, despite higher per-token prices.